perf: minimize cursor::position queries

[reubeno]

Skip redundant cursor::position queries in the repaint hot path

Reedline issues a CSI 6n / DSR (cursor::position()) on every paint to check whether prompt_start_row has drifted from the terminal's real cursor row. That's a synchronous round-trip; over high-latency transports like SSH it dominates prompt-redraw latency.

Within a read_line cycle reedline is the sole writer to the terminal and updates prompt_start_row itself as its rendering scrolls — so the drift check is unnecessary in the steady state.

A new prompt_start_row_stale: bool on Painter tracks freshness:

• Cleared after the cursor query in initialize_prompt_position.
• Set by paths that yield the tty or write content the painter doesn't model (handle_resize, print_external_message, paint_line, open_editor).
• Re-cleared on the next paint by a confirmed-no-drift query.

The hot-path drift check skips the DSR when the cache is fresh. The change is wholly inside reedline, so the improvement carries to any consumer (e.g., Nushell, brush, custom REPLs) without changes on their side.

Notes

I know that this is an area that may be challenging to test/validate. I'm open to any suggestions/feedback on how to ensure that this change does not regress existing functionality, while still improving the performance.

Full disclosure: as you can probably tell, I used Claude to identify and develop changes -- but I was heavily steering, and ultimately rewrote the code several times to address concerns that I had myself over maintainability challenges. I'm also placing this in draft first, with the hope to get early feedback on the approach and the maintainers' openness to considering this change. Any and all feedback is very much welcome, please 😄

Tradeoffs

• Each invalidating event trades the unconditional per-paint DSR for one DSR on the next paint; that's strictly better than the current per-paint cost, but not free on workloads heavy in external-printer batches or resizes.
• Terminals where cursor::position() fails see the same behavior as before: the drift check still runs and fails, with no optimization gain (and no regression).

Measurements

Nushell 0.113.0 (HEAD) built against this branch's reedline vs against reedline main. PTY harness (portable-pty), 80×24, scripted 80 ms input pacing, 5 Enter presses per run, --no-config-file --no-history. Three samples per condition; the harness is deterministic and produced identical DSR counts across samples within each condition.

DSR counts:

reedline	total DSRs (5 Enters)	per Enter
main	19	3.80
this branch	6	1.20

Same harness with --simulated-rtt-ms 125 (each CSI 6n reply delayed 125 ms — comparable to a typical cross-region SSH round trip):

reedline	total wall-clock (5 Enters)	per Enter
main	~2150 ms	~430 ms
this branch	~640 ms	~127 ms

The remaining ~1 DSR/cycle on this branch is the unavoidable one in initialize_prompt_position; reedline genuinely doesn't know where the host left the cursor before read_line is called.

brush, a smaller reedline-based shell, shows the same shape (3.0 → 1.0 DSRs/cycle) — confirming the win is at the reedline layer, not shell-specific.

[Bahex]

Suggested change

	prompt_start_row: u16,
	prompt_start_row_stale: bool,
	prompt_start_row: Option<u16>,

Using Option instead of two separate values that are semantically linked would be better. Having None instead of checking a _stale field ensures it's not possible to accidentally use a stale value.

edit: Looks like there is at least one place where the stale value can be useful. In that case a custom enum would be better. Maybe something like:

enum CachedValue<T> {
    Clear,
    Stale(T),
    Fresh(T),
}

[reubeno]

Thanks for the feedback, @Bahex. I'd considered an enum but wanted to keep the change simpler; since you're also thinking it would be an improvement, I've pushed changes to combine the fields.

I went with a custom enum so I could have some convenience helpers implemented on it, but I'm also open to a generic CachedValue<T> like you suggest above if that's what you'd recommend/prefer.

If you have time to look at the updated version, please let me know what you think.